Task Constraint-Guided Inverse Reinforcement Learning (TC-IRL) in Human-Robot Collaborative Assembly

Collaborative robots transit from the traditional robot-in-a-cell scenarios to a human-robot-shared workspace. This demands robots to better understand their human partners and then assist them. Existing robot learning from demonstration work mainly focuses on enabling robots to repeat human demonstrated tasks alone and usually require significant training efforts but have limited scalability to new tasks. This paper proposes a new task constraint-guided inverse reinforcement learning (TC-IRL) approach to learn assembly tasks from human demonstrations with significantly reduced state and action space (leading to less training data requirement) and computational efforts (landing to better real-time performance) than the conventional IRL. The TC-IRL is also extended to new geometric-scaled tasks to generate robot assistance to human in collaborative assembly. The proposed approaches are validated and evaluated through human-robot collaborative assembly experiments.